如何实现点云体素化及由密集特征得到二维伪图像特征

点云体素化

    def preprocess(self, lidar):
        #[N,4]
        # This func cluster the points in the same voxel.
        # shuffling the points
        np.random.shuffle(lidar)
        #把每个点分配到体素里，点坐标减去起始，除以每个体素的大小
        voxel_coords = ((lidar[:, :3] - np.array([self.xrange[0], self.yrange[0], self.zrange[0]])) / (
                        self.vw, self.vh, self.vd)).astype(np.int32)

        # convert to  (D, H, W)  [N,3]
        voxel_coords = voxel_coords[:,[2,1,0]]
        #np.unique该函数是去除数组中的重复数字，并进行排序之后输出。axis=0（返回二维数组的唯一行）return_inverse返回旧列表里面的值在新列表里的索引
        #return_counts返回新列表里面的行在旧列表里面的个数。
        voxel_coords, inv_ind, voxel_counts = np.unique(voxel_coords, axis=0, \
                                                  return_inverse=True, return_counts=True)
        #[N1,3],[N,1],[N1,1]N1表示非空体素的个数，voxel_counts统计每个非空体素里面的点个数

        voxel_features = []

        for i in range(len(voxel_coords)):
            #self.T每个体素里面最多点数
            voxel = np.zeros((self.T, 7), dtype=np.float32)
            #得到这个体素里面的点
            pts = lidar[inv_ind == i]
            if voxel_counts[i] > self.T:
                pts = pts[:self.T, :]
                voxel_counts[i] = self.T
            # 进行数据扩充，np.concatenate不增加维度对两向量相加
            voxel[:pts.shape[0], :] = np.concatenate((pts, pts[:, :3] - np.mean(pts[:, :3], 0)), axis=1)
            voxel_features.append(voxel)
        #voxel_features[N1,35,7],voxel_coords[N1,3]
        return np.array(voxel_features), voxel_coords

密集特征得到二维伪图像特征

def voxel_indexing(self, sparse_features, coords):
    #sparse_features[B,N1,C]  coords[N1,4]
    dim = sparse_features.shape[-1]

    dense_feature = torch.zeros(cfg.N, cfg.D, cfg.H, cfg.W, dim).to(cfg.device)

    dense_feature[coords[:,0], coords[:,1], coords[:,2], coords[:,3], :]= sparse_features

    return dense_feature.permute(0, 4, 1, 2, 3)